A sequential recommendation method based on wavelet hybrid expert framework

Abstract

The application relates to the technical field of sequence recommendation and AI intelligence, in particular to a sequence recommendation method based on a wavelet mixed expert framework. The method comprises the following steps: obtaining a historical interaction sequence of a user and converting the historical interaction sequence into an embedded representation; performing a discrete wavelet transform on the embedded representation, and decomposing the embedded representation into low-frequency approximation coefficients and high-frequency detail coefficients; for the low-frequency approximation coefficients, directly retaining a trend expert representing long-term preferences of the user; for the high-frequency detail coefficients, generating time refinement features through a learnable weight matrix element-wise multiplied with the high-frequency detail coefficients; obtaining local experts representing instantaneous interests of the user through a context-aware router; performing an inverse discrete wavelet transform, and reconstructing into an enhanced embedded representation; and adopting a multi-task learning strategy to jointly train a generated recommendation model. The method provided by the application improves the accuracy and robustness of a recommendation result.

Description

Technical Field

[0001] This application relates to the fields of sequence recommendation and AI intelligent technology, and in particular to a sequence recommendation method based on a wavelet hybrid expert framework. Background Technology

[0002] In the field of sequence recommendation, Transformer-based models (such as SASRec and BERT4Rec) have become the mainstream approach. These models capture dependencies in users' historical interaction sequences through self-attention mechanisms, achieving good recommendation results. However, real-world user behavior sequences often contain random noise such as accidental clicks, and user interests drift non-stationarily over time, making it difficult for existing models to effectively distinguish between stable preferences and transient intentions.

[0003] To overcome the aforementioned problems, frequency domain analysis-based sequence recommendation methods have emerged. These methods utilize the Discrete Fourier Transform (DFT) to transform user interaction sequences from the time domain to the frequency domain, separating long-term trends from short-term fluctuations by analyzing spectral components: the low-frequency component corresponds to asymptotically stable interests, while the high-frequency component reflects transient changes. Frequency domain methods, to some extent, improve the model's robustness to noise.

[0004] However, the global sinusoidal basis functions used in Fourier transform have inherent limitations: First, the transformation process compresses time-series information into the frequency domain, sacrificing the contextual relationships in the time dimension. This makes it impossible for the model to distinguish behaviors with the same frequency characteristics but occurring at different points in time (for example, it may be difficult to distinguish between a user's outdated interests and current intentions in the frequency domain), thus affecting the accurate inference of immediate interests. Second, Fourier transform is a single-resolution analysis, which makes it difficult to resolve signal structures at multiple scales. This makes it impossible for the model to effectively distinguish meaningful short-term interest mutations and random high-frequency noise. It often simply attenuates all high-frequency components, which may lead to the loss of key interest drift information. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application aims to provide a sequence recommendation method based on a wavelet hybrid expert framework, thereby improving the accuracy and robustness of recommendation results in the time dimension.

[0006] To achieve the above objectives, this application provides a sequence recommendation method based on a wavelet hybrid expert framework, comprising: Obtain the user's historical interaction sequence and convert it into an embedded representation; Perform a discrete wavelet transform on the embedded representation to decompose it into low-frequency approximation coefficients that retain time information and high-frequency detail coefficients that have multi-resolution characteristics; For the low-frequency approximation coefficients, they are directly retained as trend experts representing users' long-term preferences; for the high-frequency detail coefficients, time-refined features are generated through a learnable weight matrix that is multiplied element-wise with the high-frequency detail coefficients. The context-aware router dynamically assesses the importance of each frequency band based on the global context and uses the importance to modulate the time-refined features to obtain local experts that characterize the user's instantaneous interests. The trend expert and the local expert are reconstructed into an enhanced embedded representation by performing inverse discrete wavelet transform on the trend expert and the local expert. A multi-task learning strategy is adopted, which combines recommendation loss, contrastive learning loss and spectral balance regularization to jointly train and generate a recommendation model.

[0007] Furthermore, the specific steps of performing a discrete wavelet transform on the embedded representation to decompose it into low-frequency approximation coefficients that retain time information and high-frequency detail coefficients with multi-resolution characteristics are as follows: ; ; in An embedded representation of the user's historical interaction sequence. For the first Layer input, For discrete wavelet transform, For the first The low-frequency approximation coefficients of the layer, To decompose the hierarchy, For the first The high-frequency detail coefficients of the j-th decomposition level of the layer.

[0008] Furthermore, the specific steps for generating time-refined features using a learnable weight matrix that is element-wise multiplied with the high-frequency detail coefficients are as follows: ; in, For the first The high-frequency detail coefficients of the j-th decomposition level. For the first The time feature weights of the j-th decomposition level. For the first The time refinement feature of the j-th decomposition level.

[0009] Furthermore, the specific steps of dynamically evaluating the importance of each frequency band based on the global context using a context-aware router, and modulating the time-refined features using the importance to obtain a local expert representing the user's instantaneous interest, are as follows: ; ; ; ; in, For the first The context representation of a layer, This is an averaging operation along the sequence length dimension. For the first Layer gate weight set, Let the weights be the values ​​of the j-th decomposition level. For the first The gating weights of the j-th decomposition level. For local experts.

[0010] Furthermore, the specific steps for performing inverse discrete wavelet transform on the trend expert and local expert to reconstruct the enhanced embedded representation are as follows: ; ; in, For inverse discrete wavelet transform, For the first Layered augmented embedding representation, For layer normalization, This is a random deactivation function.

[0011] Furthermore, the method also includes: The enhanced embedding representation is obtained by performing a nonlinear transformation on the point-forward network.

[0012] Furthermore, the specific steps for performing a nonlinear transformation on the enhanced embedding representation using a point-forward network to obtain the transformed enhanced embedding representation are as follows: ; ; in, To enhance the embedding representation of the transformation, and These are learnable parameters. for For the first The input of the layer.

[0013] Furthermore, the specific steps for jointly training and generating the recommendation model using a multi-task learning strategy, combining recommendation loss, contrastive learning loss, and spectral balance regularization, are as follows: ; ; ; ; ; ; ; in, The item embedding matrix shares weights with the input embedding representation. For historical interaction sequence Enhanced embedding representation in layer L For users Historical interaction sequence, To predict the score matrix for clicking the next item, To indicate whether the i-th item will be clicked next, use 1 if clicked, and 0 if not. Let i be the predicted click score for the i-th item. For the total number of items, To recommend losses, and These are unsupervised views and supervised enhanced views, respectively. The final sequence representation, and These are the original view and the unsupervised augmented view obtained by encoding the same input sequence twice under different Dropout masks, respectively. To and Supervised positive sample sequences that share the same target item. To compare the learning regularization loss, For negative samples, This represents the dot product operation. For temperature coefficient, Representing the The average gating probability of each decomposition level band over the entire batch. For spectral balance regularization, and It's a hyperparameter.

[0014] To achieve the above objectives, the electronic device provided in this application includes: processor; A memory having stored thereon one or more computer program instructions that run on the processor; When the processor executes the computer program instructions, it performs the sequence recommendation method based on the wavelet hybrid expert framework as described above.

[0015] To achieve the above objectives, the present application provides a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, perform the steps of the sequence recommendation method based on a wavelet hybrid expert framework as described above.

[0016] This application presents a sequence recommendation method based on a wavelet hybrid expert framework. By introducing discrete wavelet transform to decompose user interaction sequences at multiple scales, it preserves temporal context information, overcomes the time agnosticity of Fourier transform, and can accurately distinguish between "past interest bursts" and "current immediate intentions," thereby improving the accuracy of recommendation results in the time dimension.

[0017] This application presents a sequence recommendation method based on a wavelet hybrid expert framework. It employs an asymmetric expert strategy to differentiate between low-frequency and high-frequency components and combines a context-aware router to dynamically modulate high-frequency detail coefficients. This enables the model to automatically suppress high-frequency noise (assigning low weights) during periods of stable user behavior and activate corresponding high-frequency details (assigning high weights) when interest mutations occur. This allows for the accurate capture of instantaneous and meaningful interest changes, significantly improving the robustness and response speed of the recommendation.

[0018] This application presents a sequence recommendation method based on a wavelet hybrid expert framework. It enhances the robustness of sequence representation to noise and alleviates the data sparsity problem by constructing a comparative learning loss based on unsupervised augmented views based on Dropout and supervised positive samples based on the same target item. On the other hand, spectral balance regularization prevents the context-aware router from collapsing to a few high-frequency bands by penalizing the difference between the gating weight distribution and the uniform distribution, ensuring that all high-frequency experts are used in a balanced manner. This avoids the uneven use of experts and training instability common in hybrid expert architectures.

[0019] Other features and advantages of this application will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing this application. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the present application and form part of the specification. Together with the embodiments of the present application, they serve to explain the present application but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the sequence recommendation method based on a wavelet hybrid expert framework proposed in this application. Figure 2 This is a schematic diagram of the cross-domain recommendation model in Embodiment 1 of this application; Figure 3 This is a schematic diagram of the test dataset content of Embodiment 1 of this application; Figure 4This is a schematic diagram illustrating the recommendation performance of each recommendation model; Figure 5 This is a schematic diagram of the parameters for each recommendation model. Detailed Implementation

[0021] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While some embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this application. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0022] It should be understood that the steps described in the method embodiments of this application may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this application is not limited in this respect.

[0023] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below.

[0024] It should be noted that the terms "one" and "multiple" used in this application are illustrative rather than restrictive, and those skilled in the art should understand that, unless explicitly stated otherwise in the context, they should be understood as "one or more". "Multiple" should be understood as two or more.

[0025] Terminology Explanation: Sequence recommendation, a significant branch of recommender systems, aims to predict a user's next likely interaction based on the chronological order of their historical interactions. Unlike traditional static recommendation methods such as collaborative filtering or matrix factorization, sequence recommendation fully considers the dynamic evolution of user behavior—user interests are not static but evolve over time. For example, after browsing technology products for a period, a user might suddenly become interested in gift items due to an approaching holiday. Sequence recommendation models need to capture this evolutionary pattern of interest from the user's time-ordered interaction sequences (such as clicks, purchases, and favorites) and predict the next most likely item to be selected. This technology has wide applications in e-commerce, video streaming, and music recommendation, significantly improving the timeliness and personalization of recommendations.

[0026] Discrete wavelet transform (DWT) is a signal processing technique used to decompose discrete time series into a set of different frequency components while preserving temporal information. Unlike Fourier transform, wavelet transform uses finite-length, decaying wavelet basis functions (rather than infinitely long sine waves), thus enabling the analysis of frequency components while simultaneously locating their position on the time axis—a characteristic known as time-frequency localization. DWT achieves multi-resolution analysis through layer-by-layer decomposition: each layer passes the signal through a low-pass filter to obtain "approximation coefficients" (representing low-frequency trends), through a high-pass filter to obtain "detail coefficients" (representing high-frequency details), and then downsamples. After multiple layers of decomposition, the original signal is decomposed into a final low-frequency approximation coefficient and multiple high-frequency detail coefficients at different resolutions. In this invention, DWT is used to decompose user behavior sequences into low-frequency components representing long-term preferences and multi-resolution high-frequency components representing short-term fluctuations, laying the foundation for subsequent differentiated processing.

[0027] The inverse discrete wavelet transform (IDWT) is the reverse process of the discrete wavelet transform, used to reconstruct the original time series from the coefficients after wavelet decomposition. Its basic principle is as follows: the low-frequency approximation coefficients and high-frequency detail coefficients obtained from the decomposition are upsampled, then filtered through reconstruction filters (low-pass and high-pass reconstruction filters), and the filtered results are summed to obtain the reconstructed signal of the next layer. Through layer-by-layer reconstruction, the original signal can be perfectly recovered from the decomposition coefficients of all levels (without information loss under ideal conditions). In this invention, IDWT plays a crucial role in "frequency domain to time domain conversion": after differential processing and dynamic modulation of trend experts (low frequency) and local experts (high frequency) in the frequency domain, these processed frequency domain components need to be reconstructed back to the time domain representation using IDWT for subsequent nonlinear transformation and final recommendation prediction.

[0028] For example, given a length of discrete sequence Discrete wavelet transform (DWT) utilizes the Mallat algorithm to decompose a discrete sequence into a set of high-frequency and low-frequency sub-signals through hierarchical decomposition. Level decomposition is defined as: ; ; index Achieved step size of downsampling, This is the filter length. Indicates after the first Level low-pass filter The approximation coefficients after that contain the low-frequency components of the signal. When At that time, we set . Indicates after the first High-pass filter The detailed coefficients after wavelet decomposition contain the high-frequency components of the signal. Through wavelet decomposition, the Discrete Wavelet Transform (DWT) can locate transient components in a time-domain signal, thus enabling the processing and analysis of non-stationary signals. The effectiveness of this decomposition depends on the vanishing moments. In mathematics, it has " "A wavelet with vanishing moments of order 1 implies that its high-pass filter H is the highest" Polynomials of order "[number]" are orthogonal. This can be explicitly expressed as a moment condition: ; This orthogonality ensures that the smooth polynomial tendency is suppressed in the detailed coefficients. Therefore, higher... The value can produce a sparser high-frequency representation, effectively decoupling transient signals from long-term trends.

[0029] Furthermore, the decomposed high-frequency and low-frequency sub-signals can be perfectly reconstructed from the original signal using the inverse discrete wavelet transform (IDWT). This process involves iterative upsampling and filtering operations to reconstruct the signal in stages. ; in and It is a reconstruction filter; the forward DWT converts the sequence behavior into high / low frequency sub-signals and uses... Indicated by IDWT (using (This indicates that) the decomposed sub-signals are reconstructed into the original signal.

[0030] Hybrid experts are a machine learning architecture that combines multiple "expert" networks and a "gating" network to achieve complex function mappings. Its core idea is "divide and conquer": the input space is divided into multiple regions, with each expert network specializing in processing a specific type of input pattern; the gating network dynamically calculates the weights of each expert based on the input, and the final output is a weighted sum of the outputs of all experts. Standard MoEs typically employ multiple homogeneous experts with identical structures to expand model capacity. In this invention, hybrid experts represent a heterogeneous expert strategy—the experts are no longer homogeneous but are specifically designed based on frequency domain characteristics: "trend experts" that directly retain low-frequency components and dynamically modulated high-frequency "local experts" are responsible for modeling long-term stable preferences and short-term transient interests, respectively, achieving a more refined behavioral decoupling than homogeneous experts.

[0031] For example, the Mixture of Experts (MoE) module maintains a set of... A network of experts It employs a learnable gating network. To dynamically allocate computing resources. For a given input Final output Calculate the weighted sum of all expert outputs: ; Where, represents the assignment to the first The gating weight of each expert, and Existing MoE methods mainly focus on using multiple homogeneous experts (i.e., experts with the same architecture) to expand the model capacity.

[0032] Contrastive learning, a self-supervised representation learning paradigm, aims to shorten the distance between "positive sample pairs" and widen the distance between "negative sample pairs" in the feature space. In sequence recommendation, contrastive learning can be used to enhance the robustness of representations to noise and alleviate data sparsity problems. This invention employs a Dropout-based enhancement strategy to construct positive sample pairs: the same input sequence is passed through an encoder twice under two different Dropout masks, resulting in two different views that share the same semantic information but have different random noise patterns, forming unsupervised positive sample pairs. Simultaneously, supervised positive samples are introduced—sequences in the same training batch that have the same next interactive item are considered positive sample pairs because they share the same prediction target. Through InfoNCE loss function optimization, the model can learn noise-insensitive and discriminative sequence representations.

[0033] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0034] Example 1 One embodiment of this application provides a sequence recommendation method based on a wavelet hybrid expert framework. Figure 1 This is a flowchart illustrating the sequence recommendation method based on a wavelet hybrid expert framework proposed in this application. Figure 2 This is a schematic diagram of the cross-domain recommendation model in Embodiment 1 of this application; the following will refer to... Figures 1-2 This application provides a detailed description of the sequence recommendation method based on a wavelet hybrid expert framework.

[0035] Step S101: Obtain the user's historical interaction sequence and convert it into an embedded representation.

[0036] In this application embodiment, for a user's historical interaction sequence We use an item embedding matrix To define the embedded representation of historical interaction sequences , corresponding to Figure 2 The embedding layer in the code. We use position embedding. To add additional location information while preserving the original dimensionality of the item embeddings, we perform layer normalization and dropout operations to stabilize the training process. Therefore, the generated embedding representations... as follows: .

[0037] Step S102: Perform discrete wavelet transform on the embedded representation to decompose the embedded representation into low-frequency approximation coefficients that retain time information and high-frequency detail coefficients with multi-resolution characteristics; In this application's implementation, to address the limitations of global Fourier transform in capturing local details, we propose a mechanism inspired by multiresolution analysis. Its core concept is "wavelet localization and gated selection." It embeds the input... The model is decomposed into different frequency subbands that retain time information. Then, based on the context, the dominant frequency band containing key user behavior patterns is dynamically identified and enhanced, while irrelevant noise frequency bands are suppressed, enabling the model to distinguish between long-term stable trends and short-term transient patterns.

[0038] like Figure 2 As shown, in the recommendation model framework of this application, there are L stacked processing layers (i.e., the purple box on the left in the figure). Each processing layer includes a frequency mixture of experts and a point-forward neural network. The frequency mixture of experts first performs wavelet-based multi-scale decomposition, performs one-dimensional discrete wavelet transform (1DDWT) along the time dimension of E^u, projects the sequence to the time-frequency domain, and explicitly decouples it into a series of orthogonal frequency sub-bands.

[0039] We employ Symlets wavelets because their near-linear phase and symmetry minimize phase distortion compared to asymmetric wavelets (such as Daubechies). To handle boundary effects, we use a zero-filling strategy. The decomposition process is defined as follows: ; ; in An embedded representation of the user's historical interaction sequence. For the first Layer input, For discrete wavelet transform, For the first The low-frequency approximation coefficients of the layer, To decompose the hierarchy, For the first The high-frequency detail coefficients of the j-th decomposition level of the layer.

[0040] Step S103: For the low-frequency approximation coefficients, directly retain them as trend experts representing users' long-term preferences; for the high-frequency detail coefficients, generate time-refined features through a learnable weight matrix that is multiplied element-wise with the high-frequency detail coefficients. In this application embodiment, in order to explicitly model the decoupled information, a differentiated processing strategy is adopted for low-frequency and high-frequency components; Due to approximation coefficients These represent stable long-term user preferences, so they are directly retained as robust anchors for recommendation stability, serving as trend experts; For detail factor We introduce temporal feature modulation. We define a learnable weight matrix. It acts as a time filter, learning the weights at specific locations through element-wise multiplication to highlight or suppress details at specific time steps. ; in, For the first The high-frequency detail coefficients of the j-th decomposition level. For the first The time feature weights of the j-th decomposition level. For the first The time refinement feature of the j-th decomposition level.

[0041] Step S104: The context-aware router dynamically evaluates the importance of each frequency band based on the global context, and modulates the time-refined features using the importance to obtain a local expert characterizing the user's instantaneous interest. Then, to distinguish meaningful bursts of interest from random noise, we designed a context-aware gating mechanism. It dynamically evaluates the importance of each frequency band based on the global context. Specifically, a two-layer MLP with a ReLU activation function generates a Sigmoid distribution over all detailed features: ; ; ; ; in This represents the importance of each frequency band. For the first The context representation of a layer, This is an averaging operation along the sequence length dimension. For the first Layer gate weight set, Let the weights be the values ​​of the j-th decomposition level. For the first The gating weights of the j-th decomposition level. For local experts.

[0042] High weighting indicates that the frequency band contains key signals of the current intent, while low weighting means that the frequency band is dominated by noise, and the output... Modulating time-refined features as a global gating signal .

[0043] This hierarchical design achieves a dual optimization of temporal accuracy and semantic relevance.

[0044] Step S105: Perform inverse discrete wavelet transform on the trend expert and the local expert to reconstruct an enhanced embedded representation; Finally, we use inverse-DWT to reconstruct all modulated components back to the time domain and apply residual connections and normalization: ; ; in, For inverse discrete wavelet transform, For the first Layered augmented embedding representation, For layer normalization, This is a random deactivation function.

[0045] In this embodiment, to endow the model with nonlinear characteristics across different dimensions in the time domain, we add a feedforward network after each feature mixer. This network consists of an MLP with the GELU activation function. The pointwise feedforward neural network (FFN) process is defined as follows: to prevent overfitting, a Dropout layer is added above each hidden layer, and the output is again processed using a residual connection structure. Execution layer normalization procedure: ; ; in, To enhance the embedding representation of the transformation, and These are learnable parameters. for For the first The input of the layer.

[0046] Step S106: Employ a multi-task learning strategy, combining recommendation loss, contrastive learning loss, and spectral balance regularization to jointly train and generate a recommendation model; In this embodiment, to enhance the robustness of the model and prevent expert collapse, we employ a multi-task learning strategy, combining recommendation loss, contrastive learning loss, and spectral balance regularization for joint training, as shown in the following formula: ; ; ; ; ; ; ; in, The item embedding matrix shares weights with the input embedding representation. For historical interaction sequence Enhanced embedding representation in layer L For users Historical interaction sequence, To predict the score matrix for clicking the next item, To indicate whether the i-th item will be clicked next, use 1 if clicked, and 0 if not. Let i be the predicted click score for the i-th item. For the total number of items, To recommend losses, and These are unsupervised views and supervised enhanced views, respectively. The final sequence representation, and These are the original view and the unsupervised augmented view obtained by encoding the same input sequence twice under different Dropout masks, respectively. To and Supervised positive sample sequences that share the same target item. For negative samples, To compare the learning regularization loss, This represents the dot product operation. For temperature coefficient, Representing the The average gating probability of each decomposition level band over the entire batch. For spectral balance regularization, and It's a hyperparameter.

[0047] Regarding the recommendation loss, the main training objective is to minimize the cross-entropy loss between the predicted probability distribution ŷ and the one-hot vector y of the true label. The loss function prompts the model to increase the predicted probability of the target item while decreasing the probability of other non-target items.

[0048] Regarding contrastive learning, instead of employing typical data-level augmentations (such as item cropping, masking, and reordering) that might compromise semantic integrity, a Dropout-based augmentation strategy was used. Specifically, the same input sequence was passed twice through a multi-expert frequency domain encoder under different Dropout masks, resulting in two views: an unsupervised augmented view. and original view Furthermore, in order to utilize the monitoring signals between sequences, we introduced a method related to... A sequence that shares the same target item (Next Item). As a supervised positive sample, its representation is denoted as For a training batch, the contrastive learning regularization loss is used. Calculate using the formula above.

[0049] Regarding spectral balance regularization, in this embodiment, a load balancing regularization term is introduced as described above to prevent the soft gating mechanism (Context-Aware Router) from collapsing to a very small number of frequency bands (i.e., preventing the model from focusing only on a few specific frequencies while ignoring others). By penalizing cases where the gating weights deviate from a uniform distribution, the model is encouraged to focus on all frequencies. Maintain uniform attention across every detail frequency band.

[0050] like Figure 2 As shown, the model's prediction layer selects the representation of the last interacted item. To calculate the predicted scores for all candidate items: ; in It is an item embedding matrix that shares weights with the input embedding.

[0051] In this application's implementation, experiments were conducted on four publicly available datasets collected from real-world platforms to comprehensively evaluate the trained recommendation model (WEARec). See [link to relevant documentation]. Figure 3 , Figure 3 This is a schematic diagram of the test dataset content of Embodiment 1 of this application, as shown below. Figure 3As shown, the datasets include Beauty and Sports from Amazon, ML-1M, and LastFM.

[0052] See Figure 4 , Figure 4 This diagram illustrates the recommendation performance of each recommendation model. This embodiment employs a leave-one-out strategy to partition each user's item sequence, ranking the predicted scores across the entire item set without using negative sampling. Performance is evaluated using the hit rate (HR@K) and normalized discount cumulative gain (NDCG@K, NG@K) at K values ​​of 10 and 20, respectively. Figure 4 As shown, the recommendation model (WEARec) trained in this application improves recommendation performance by nearly 1%-5% compared to traditional recommendation models.

[0053] See Figure 5 , Figure 5 These are the optimal parameters for the recommendation model (WEARec) on the four datasets.

[0054] Example 2 In this embodiment, an electronic device is also provided, including a processor and a memory. The memory is used to store non-transitory computer-readable instructions. The processor is used to execute the non-transitory computer-readable instructions, which, when executed by the processor, can perform one or more steps of the sequence recommendation method based on the wavelet hybrid expert framework described above. The memory and the processor can be interconnected via a bus system and / or other forms of connection mechanisms.

[0055] For example, the processor can be a central processing unit (CPU), a digital signal processor (DSP), or other processing units with data processing and / or program execution capabilities, such as a field-programmable gate array (FPGA); for example, the central processing unit (CPU) can be an x86 or ARM architecture.

[0056] For example, the memory may include any combination of one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, erasable programmable read-only memory (EPROM), portable compact disc read-only memory (CD-ROM), USB storage, flash memory, etc. One or more computer program modules may be stored on the computer-readable storage medium, and the processor may run one or more computer program modules to implement various functions of the electronic device. Various application programs and various data, as well as various data used and / or generated by the application programs, may also be stored in the computer-readable storage medium.

[0057] It should be noted that the specific functions and technical effects of the electronic devices in the embodiments of this application can be referred to the description of the sequence recommendation method based on wavelet hybrid expert framework above, and will not be repeated here.

[0058] Example 3 In this embodiment, a computer-readable storage medium is also provided for storing non-transitory computer-readable instructions. For example, when the non-transitory computer-readable instructions are executed by a computer, one or more steps in the sequence recommendation method based on the wavelet hybrid expert framework described above can be performed.

[0059] For example, this storage medium can be used in the aforementioned electronic device. For example, the storage medium can be the memory in the electronic device of Embodiment 2. For example, relevant descriptions of the storage medium can be found in the corresponding description of the memory in the electronic device of Embodiment 2, and will not be repeated here.

[0060] It should be noted that the storage medium (computer-readable medium) mentioned above in this application may be a computer-readable signal medium or a non-transitory computer-readable storage medium, or any combination of the two. A non-transitory computer-readable storage medium may be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a non-transitory computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0061] In this application, a non-transitory computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a non-transitory computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination thereof.

[0062] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0063] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include, but are not limited to, object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server.

[0064] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer program instructions.

[0065] The units described in the embodiments of this application can be implemented in software or hardware. The names of the units are not, in some cases, limiting the scope of the unit itself.

[0066] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), etc.

[0067] The above description is merely a partial embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

[0068] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in sequential order. Multitasking and parallel processing may be advantageous in certain environments. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this application. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0069] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A sequence recommendation method based on a wavelet hybrid expert framework, comprising: Obtain the user's historical interaction sequence and convert it into an embedded representation; Perform a discrete wavelet transform on the embedded representation to decompose it into low-frequency approximation coefficients that retain time information and high-frequency detail coefficients that have multi-resolution characteristics; the high-frequency detail coefficients can locate transient components in the time-domain signal. For the low-frequency approximation coefficients, they are directly retained as trend experts representing users' long-term preferences; for the high-frequency detail coefficients, time-refined features are generated through a learnable weight matrix that is multiplied element-wise with the high-frequency detail coefficients. By using a context-aware router, the importance of each frequency band is dynamically evaluated based on the global context, and the temporally refined features are modulated using the importance to obtain a local expert representing the user's instantaneous interest. The specific steps are as follows: ； ； ； ； in, For the first The context representation of a layer, This is an averaging operation along the sequence length dimension. For the first Layer gate weight set, Let the weights be the values ​​of the j-th decomposition level. For the first The gating weights of the j-th decomposition level. The trend expert and the local expert are respectively responsible for modeling long-term stable preferences and short-term transient interests; The trend expert and the local expert are subjected to inverse discrete wavelet transform to reconstruct an enhanced embedded representation. A multi-task learning strategy is adopted, which combines recommendation loss, contrastive learning loss and spectral balance regularization to jointly train and generate a recommendation model.

2. The sequence recommendation method based on a wavelet hybrid expert framework according to claim 1, characterized in that, The specific steps of performing discrete wavelet transform on the embedded representation to decompose it into low-frequency approximation coefficients that retain time information and high-frequency detail coefficients with multi-resolution characteristics are as follows: ； ； in An embedded representation of the user's historical interaction sequence. For the first Layer input, For discrete wavelet transform, For the first The low-frequency approximation coefficients of the layer, To decompose the hierarchy, For the first The high-frequency detail coefficients of the j-th decomposition level of the layer.

3. The sequence recommendation method based on a wavelet hybrid expert framework according to claim 2, characterized in that, The specific steps for generating time-refined features using a learnable weight matrix that is element-wise multiplied with the high-frequency detail coefficients are as follows: ； in, For the first The high-frequency detail coefficients of the j-th decomposition level. For the first The time feature weights of the j-th decomposition level. For the first The time refinement feature of the j-th decomposition level.

4. The sequence recommendation method based on a wavelet hybrid expert framework according to claim 3, characterized in that, The specific steps for performing inverse discrete wavelet transform on the trend expert and the local expert to reconstruct the enhanced embedded representation are as follows: ； ； in, For inverse discrete wavelet transform, For the first Layered augmented embedding representation, For layer normalization, This is a random deactivation function.

5. The sequence recommendation method based on a wavelet hybrid expert framework according to claim 4, characterized in that, The method further includes: The enhanced embedding representation is obtained by performing a nonlinear transformation on the point-forward network.

6. The sequence recommendation method based on a wavelet hybrid expert framework according to claim 5, characterized in that, The specific steps for performing a nonlinear transformation on the enhanced embedding representation using a point-forward network to obtain the transformed enhanced embedding representation are as follows: ； ； in, To enhance the embedding representation of the transformation, and These are learnable parameters. for For the first The input of the layer.

7. The sequence recommendation method based on a wavelet hybrid expert framework according to claim 6, characterized in that, The specific steps for jointly training and generating the recommendation model using a multi-task learning strategy, combining recommendation loss, contrastive learning loss, and spectral balance regularization, are as follows: ； ； ； ； ； ； ； in, The item embedding matrix shares weights with the input embedding representation. For historical interaction sequence Enhanced embedding representation in layer L For users Historical interaction sequence, To predict the score matrix for clicking the next item, To indicate whether the i-th item will be clicked next, use 1 if clicked, and 0 if not. Let i be the predicted click score for the i-th item. For the total number of items, To recommend losses, and These are unsupervised views and supervised enhanced views, respectively. The final sequence representation, and These are the original view and the unsupervised augmented view obtained by encoding the same input sequence twice under different Dropout masks, respectively. To and Supervised positive sample sequences that share the same target item. For negative samples, To compare the learning regularization loss, This represents the dot product operation. For temperature coefficient, Representing the The average gating probability of each decomposition level frequency band over the entire batch. For spectral balance regularization, and It's a hyperparameter.

8. An electronic device, characterized in that, include: processor; A memory having stored thereon one or more computer program instructions that run on the processor; When the processor executes the computer program instructions, it performs the sequence recommendation method based on the wavelet hybrid expert framework as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, It stores computer program instructions, which, when executed by a processor, perform the steps of the sequence recommendation method based on a wavelet hybrid expert framework as described in any one of claims 1 to 7.

Images